Not notes. Not summaries. Not a to-do list you still have to do. Nebula listens to any conversation, plans the work it implies, verifies each step is right — then does it, with the right level of human control.
"I'll send Acme the revised proposal by Friday and move the deal to Negotiation."
An execution agent, not a note-taker — judged on the work it completes, not the summary it writes.
OS-level capture on Windows and macOS. Switch on system audio, your microphone, the screen, the camera — or any combination. A video call, a phone call, a hallway chat, a screen share in an app with no API. If it's a conversation, Nebula can act on it.
A neural model proposes the actions; symbolic checks verify each one is grounded in what was actually said, complete enough to execute, and authorized — before anything runs.
Reversible internal work runs automatically; drafts wait for your review; anything external and irreversible needs your approval. You set how much autonomy Nebula gets, and move the line as trust builds.
Every conversation generates work — a proposal to send, a ticket to file, a record to update. All of it was decided out loud, by people who were paying attention. Then someone sits down afterward and does it again by hand.
That second pass is the whole cost. Not the meeting, and not remembering the meeting — the re-entry.
After the call ends
Every tool stops here
One line in a meeting. Three pieces of finished work — planned, verified, and executed while the call was still going.
"Let's get Acme the revised proposal by Friday. Someone loop in legal, and I'll bump the deal to Negotiation."
Utterance · 00:14:22 · speaker: Maya
Illustration of a single utterance and the three actions Nebula plans from it.
Recording conversations, watching screens, and doing work on your behalf is three stacked responsibilities. We designed for each one from day one.
Audio, screen and camera frames are processed on your machine, and you choose which of those to capture at all. A window and app blocklist keeps passwords and unrelated DMs out of the record.
Our engine is trained on data we own and license — your conversations are never training material. Retention is customer-controlled, so you set how long recordings live.
Every action records its source utterance, timestamp, confidence, who approved it, and what the target system returned. Evidence, not a black box.
Visible consent capture and a documented process for two-party-consent states and GDPR — not a buried checkbox in settings.
Write access is scoped per tool and per action, not granted wholesale. External and irreversible actions default to requiring your approval.
The Type II process begins as we onboard paying customers. Design partners get a documented security posture and a direct line to the team.
Bundled assistants keep you inside one app. Nebula acts across the tools your work actually lives in — one conversation, many systems.
Record any conversation — system audio, microphone, screen and camera, each switched on or off per session — with live transcription, speaker labels and a local-first archive. Works in the browser, or install the desktop app.
We'll send your workspace link and the Windows and macOS installers. The desktop app adds OS-level system-audio and full-screen capture.
The layer that turns those conversations into completed work — planning, verification, and cross-tool execution with autonomy you control. Onboarding a small group of design partners now.
Nebula is an execution agent. It captures any conversation, works out the concrete actions that conversation implies, verifies each one against what was actually said, and then carries them out across your tools — filing the ticket, drafting the email, updating the CRM record. The unit of value is completed work, not a summary.
Note-takers stop at the record: a transcript, a summary, and a list of action items a human still has to do. Nebula does the work. It plans the actions a conversation implies, verifies they are grounded in what was said, and executes them in the systems where the work actually lives. Judge a note-taker on the quality of its summary; judge Nebula on the work it completes.
CRM systems (Salesforce, HubSpot, Pipedrive), issue trackers (Jira, Linear, Asana), docs and knowledge bases (Notion, Google Docs, Confluence), comms (Slack, Gmail, Outlook), calendars, and storage (Google Drive, Dropbox). Write access is scoped per tool and per action rather than granted wholesale. See the full list.
Only the ones you allow. Reversible internal actions can run automatically, drafts wait for your review, and anything external or irreversible requires your approval by default. You set the autonomy level and move it as trust builds. Every action is logged with its source utterance, timestamp, confidence, approver, and the response from the target system.
No. Your conversations are never used to train models. Capture is local-first — audio, screen and camera frames are processed on your machine — and retention is customer-controlled. The Nebula engine is trained on data we own and license, never on customer conversations. Enterprise can run it self-hosted so transcripts never leave their network. More on data handling.
Nebula Capture is the part you can use now. It runs in the browser and as native desktop apps for Windows and macOS, recording any conversation with system audio, microphone, screen and camera as separate sources you switch on per session, plus live transcription and speaker labels. The desktop app is required for OS-level system-audio and full-screen capture. The execution agent, which plans and carries out the work, is in private beta with a small group of design partners.
The capture apps are free. Paid plans start at $18 per user per month billed annually, priced per seat plus a monthly allowance of executed actions and metered above that allowance — never priced on minutes transcribed. Bringing your own AI key means no markup on model usage. Full pricing.
Join the private beta for the execution layer. A few design-partner spots each week.
Built by the team behind Nebula's OS-level capture stack for Windows and macOS, with verification research at the core of how the agent decides what is safe to run.
